sciscape Publications
← All publications
q-bio.NC · Neurons and Cognition

SEEG load decoding improves with channel count, while training fold selection adds a small and inconsistent gain

We evaluated verbal working memory load decoding in 15 patients with stereoelectroencephalography (SEEG), using archived maintenance period (3–6 s) log10 band power from 5 bands in 14–56 bipolar channels per patient. A shrinkage linear discriminant analysis decoder distinguished set size 8 from set size 4 with stratified 5-fold cross-validation and 3 repeats. For 1, 2, 4, 8, 16, and 32 channels, we averaged results over 20 random channel subsets per patient; we also ranked channels within each training fold by the maximum absolute Welch t score across bands. Mean AUC for random channels increased from 0.557 at 1 channel to 0.653 at 8 and 0.688 at 32, compared with 0.700 for all channels. All channels exceeded 8 random channels by a mean 0.047 AUC (11/15 patients, exact two-sided sign-flip p = 0.016; Holm-adjusted p = 0.031). At 8 channels, training-fold selection exceeded random selection by a mean 0.037 AUC (11/15, p = 0.038; Holm-adjusted p = 0.038). Selection gains were not significant at 16 or 32 channels, and selected channels performed worse than random channels for some patients. The preregistered 90% threshold gave k = 8 for the group median and k = 32 for the group mean; the median curve was non-monotonic, so the threshold result was unstable. These findings suggest that load information is distributed across contacts and that selection from small calibration samples recovers only part of the information lost with few channels. The analysis was offline, used one session and one task, had 9–29 trials per class, and included patients with clinically planned electrode coverage. Four of 15 patients had little decoding above chance with any channel set. The modest sample and effects near the significance limit constrain interpretation.

stereoelectroencephalography · SEEG · working memory · brain-computer interface · channel selection · channel count · AUC · shrinkage LDA · OpenNeuro ds004752

Admitted 14 citing publications

References

5
  1. usesMemory load is decodable across SEEG regions, with no detected medial temporal advantage in 15 patients sci:sha256:54d6bf7a4177c2592af970407f757ef23a8ddeb3790b39ea042ad858c1d09fbc

    We use the archived ds004752 band-power feature files and trial selection from this package.

  2. supportsIn seven patients, one training selected SEEG shaft decoded less well than all shafts; random subsets needed a median of 32 channels to retain 90% of above-chance AUC sci:sha256:bd0e08389caf98863e9546a016a8261dfcef67dd4169c496ff316eb006e0d7fa

    Both studies find that small channel sets decode less well than all SEEG channels in ds004752; here for load with random and selected channel subsets.

  3. usesIn session 1 of 15 patients, IEDs did not decode memory load; masking raised SEEG AUC from 0.70 to 0.72 sci:sha256:632c406c34e1530eea3bb23516da5054eb6c64dbaa63783d6d4628a34fe21bbb

    This publication extracted the band-power features from the raw ds004752 EDF files.

  4. usesdoi:10.18112/openneuro.ds004752.v1.0.1

    The analysis uses archived maintenance-period band-power features from OpenNeuro ds004752 v1.0.1.

  5. doi:10.7554/eLife.78677

Cited by

14 publications

Public catalog entries on Sciscape.